[mikecsh]

Great! Let me hook you up with the latest medical AI technology. Actually, let me hook you up with all of the latest medical AIs, throw your symptoms at them and let them treat all you ailments. And add to that, all the latest medical robotics, and we'll take the surgeons and intensivists, interns, and nurses away.

Good luck with that, I'm sure you will be much happier and healthier when the current AI tries to decide whether you are acutely dying of hypercalcaemia or a cardiac tamponade and Marshalls the robots to save you.

\s

For all the faults and failings of doctors (read: humans), I find many Hacker Newsers vastly oversimplify what doctors do and vastly overinflate what AI is able to do now, or will be able to do in the next ten, twenty, or even 30 years.

It's great that we not have some AI that can diagnose a stroke better than a radiologist on a CT scan. But it can't tell you the scan also shows a space occupying lesion. Or hydroencephaly. Or an extra dural haematoma. Or any other intracranial pathology. And that's about the state of the art.

I welcome new diagnostic aids as I think most clinicians do, but I would take the fleshy system we have of human healthcare providers and suspect that will be the case for a good while yet.

Disclaimer: am doctor and engineer

[notabee]

Honestly, many of the obstacles seem to be social and procedural ones. We already have the technology to easily facilitate better communication between e.g. PCPs and specialists to better diagnose rare diseases, but it's not used well. Medical centers are still faxing everything around, doing manual data migrations, and have little in the way of technological standards. There are flatly stupid traditions such as doctors working 24+ hours at a stretch when we force truckers and airline pilots to limit their hours far below that for less mentally demanding jobs. The social structure of medicine is an ossified, dysfunctional mess and all the diagnostic aids in the world aren't going to fix that. That's why I feel that people want to replace it all, regardless of how feasible it might be right now. Fixing the logistical, economic, and procedural problems that prevent people who need to be in a scanner from getting there quickly enough will do substantially more than getting a few percentage points of extra sensitivity or specificity by letting an AI read the scans. We don't need some far future technology, we need medicine to use existing technology intelligently.

[elevenoh]

>It's great that we not have some AI that can diagnose a stroke better than a radiologist on a CT scan. But it can't tell you the scan also shows a space occupying lesion. Or hydroencephaly. Or an extra dural haematoma. Or any other intracranial pathology. And that's about the state of the art.

Why couldn't a model report all discernable co-morbidities?

I don't see any form of objective data analysis remaining optimally in the hands of humans for long.

Double checked by humans, in the short term, sure.

There's arguably zero 'creativity' in diagnostics. It'll be automated within a decade or two IMO.

[omgwtfbyobbq]

There's a lot of nuance in diagnostics, and information can be scattered about, especially with relatively rare disorders.

[mikecsh]

> Why couldn't a model report all discernable co-morbidities?

It could, but it's not there yet. I was trying to illustrate the enormous chasm which AI has yet to cross. Even if an AI model is trained to interpret all head CT pathologies, that is an absolutely minuscule part of practicing medicine.

Let me be clear that I agree that many of the functions of doctors are theoretically replicable with technology. I just have radically different view to the timescale that this will be on compared to many HN-ers who seem to equate treating patients with analysing a computer program, a comparison which is woefully inadequate.

The advances in image interpretation AIs in medicine are a bit misleading as they are literally the lowest of the low hanging fruit. There are huge amounts of data to mine with both normal results and pathological ones, and the data is already in a relatively consistent and nice format for the model to be trained on.

And yet it's 2020 and we don't even have accurate computerised diagnostics for ECG interpretation which is essentially 12 arrays of floats.

I relish the advances in tech, but the chasm between what "robo-doctors"/"AI" etc can actually achieve right now to benefit patients and what a doctor even just out of medical school can do on a day to day basis is vast. The progress in "potential doctor replacements" we have seen from the technology sector is hugely hyped but realistically has a minuscule effect on patient outcomes at present.

I understand people become very frustrated with inadequate healthcare systems and especially when mistakes are made. The go-to answer of "doctors are scumbags, they don't do anything anyway and AI will replace them in a decade" is facile and ill-informed.

Yesterday I walked past a patient who was vomiting fresh blood. She needed urgent wide bore IV access, bloods, blood transfusion, review from upper GI surgeons and head and neck surgical oncology and immediate return to theatre to open up her neck and explore what was going on to hopefully fix it. There is not the slightest hint of a technological solution to this managing this single random example of which I could have picked many thousands more.

Last week I was called to ED to review a patient who had had an industrial accident with heavy machinery and had an almost complete degloving of his arm and almost complete amputation of the same. The diagnosis is easy in this case, but which software is going to keep the man alive and try and save his arm?

---

Diagnosis may not be "creative" but as another commenter points out it is nuanced. I've commented about this before but the main challenge with tech for diagnosis is data collection. It's easy to train and run models on numerical data such as CT scans and blood results because the data collection is easy. Even then the tools we have are pretty useless at present. In my hospital we have an automatic warning if the system suspects a patient has sepsis, which is great. However, it's also often wrong. Which, were it allowed to and able to manage the patient would be a complete disaster. ECGs provide automated interpretation which are usually complete BS.

However, patients are not numbers, or programs; eliciting the information you need to make the diagnosis is the hardest part and that is significantly more challenging than a flowchart. E.g. patient found unresponsive at 3am by a canal, no further info. Where is the AI solving this right now?

Once you have a diagnosis, or at least a working diagnosis, you then need to be able to actually instigate the management for that patient. I have yet to see anything that can automatically take blood, cannulate a patient, intubate a patient, perform a ring block with local anaesthetic, run a cardiac arrest, etc.

The chasm looks small from a distance but when you get up close, it's actually really big.